MCP server
Prompts and resources
Six pre-built workflows a client can surface as slash commands, eight readable resources the model pulls only when it needs them, and the gtm-research skill.
Tools are what the server can do. Prompts and resources are what it knows: the workflows worth following and the reference material behind them. Both are free. A prompt costs nothing to list, and a resource costs nothing until it is read.
Prompts#
Pre-built starting points, which a client can surface as slash commands or a template picker. Each one sets up a whole workflow: the order to work in, which calls are free, and which caveats to read before reporting anything.
| Prompt | What it sets up | Arguments |
|---|---|---|
build-target-list | Coverage check, then search, then enrich, with the caveats read and the estimate flags honoured | icp, count |
find-lookalikes | Resolve the examples to domains first, then anchor on them | examples |
enrich-list | Batched lookups that respect the ceilings, with misses reported rather than dropped | identifiers, fields |
find-warm-intro | Who connects you to a target company, and how strong each connection actually is | target, from |
deep-research | A brief no single search answers: per-row criteria, named columns, exclusions. Checks first that a search would not do, and states the price before spending it. A run is charged only if it succeeds | brief, columns, effort |
check-before-you-promise | An honest answer about coverage, including the published gaps, before anything is promised | ask |
Two are worth calling out because of what they prevent:
find-lookalikesresolves your examples to domains before anchoring on them. A name anchor resolves silently and may resolve to the wrong company: an error that is invisible once it is a row in a ranked list.find-warm-introtakesfromas well astarget. Without your own side there is no second side, and a warm path needs one. A single company and nothing else is a person search wearing the wrong name.
Resources#
Reference the model can read on demand. Nothing here is billed and nothing is loaded into context until it is asked for, which is the point: detail that would be expensive as part of every request is cheap as a resource.
| URI | What it holds |
|---|---|
sprouts://guide | The full agent guide: the same content as For agents |
sprouts://errors | Every error code an API key or a connected MCP client can receive, with what to do about each (retry, back off, stop, or report it) and a link to the full error codes table. Codes used only by the dashboard sign-in flow are left out |
sprouts://coverage | What this deployment holds right now, including the published gaps |
sprouts://fields/{category} | The field catalogue for one entity category |
sprouts://filters/{name} | Accepted values and operators for one filter |
sprouts://filters/{category}/{name} | The same question for one document category, where the values differ |
sprouts://sorts/{category} | The fields a search can be ordered by, per category |
sprouts://feedback-tags | The vocabulary a feedback report can use, read live |
sprouts://guideis a copy of the For agents page, re-synced from it when the page changes. The server's CI checks that nobody has edited the copy by hand, but it does not compare it with this site, so for a short time after the page changes the two can differ. When they do, this site is the current version. The server rewrites the hosts in the guide to its own deployment's when it is read, so the URLs a model sees there are the ones it should call.- Hosts that render cards will also see
ui://resources in the list. Those are the card bundles themselves: served to the host, not meant to be read.
Oversized records#
Some records are large: a trend report averages around 4.5 MB. Rather than inline one of those into the conversation, the server stores the record it already charged you for and returns a resource link the client can open on demand.
Skills#
The packaged distributions (the npm server and the Claude Code plugin) bundle a gtm-research skill: a SKILL.md plus nine reference files covering:
- how to choose between the tools
- how to read the response envelope
- coverage and honesty
- errors and retries
- five worked patterns: target lists, look-alikes, enrichment, buying committees and research runs
It is prose a coding agent loads when the work calls for it: the long form of what the server's own instructions say in a paragraph.